Attention Spiking U-Net for Weather Forecasting

Due to the efficient and low-energy consumption advantages of event-driven spike computations in spiking neural networks, they are emerging as promising energy-saving alternatives to artificial neural networks. Weather data typically has high resolution and large volumes, and existing weather prediction methods based on artificial neural networks often consume significant amounts of energy when processing such extensive high-resolution data. This energy consumption limitation highlights the importance of applying spiking neural networks in this field. In this context, this research combines the high energy efficiency of spiking neural network with the simple and efficient architecture of U-Net, proposing a novel Attention Spiking U-Net (AS-UNet) designed to achieve high prediction performance with lower energy consumption in weather prediction tasks. Specifically, AS-UNet reduces energy consumption through spike-driven residual connections and a low-latency direct training strategy based on surrogate gradients, while enhancing model representational capacity and prediction accuracy by incorporating two types of attention mechanisms. The effectiveness of the proposed AS-UNet is validated on the WeatherBench benchmark dataset. This study explores the potential of spiking neural networks in the field of weather forecasting and offers an energy-efficient alternative for this domain.